LinearStandardFormCompiler sometimes says expression are nonlinear even when variables are fixed

[michaelbynum]

Summary

Sometimes, the LinearStandardFormCompiler throws an error saying that an expression is nonlinear even if the variables that would make it nonlinear are fixed.

Steps to reproduce the issue

In order to get a simple example that illustrates the problem, I created an instance of LinearRepnVisitor the same way that LinearStandardFormCompiler makes one. See this line and the few lines before it.

In [1]: import pyomo.environ as pe
   ...: from pyomo.repn.linear import LinearRepnVisitor
   ...: from pyomo.repn.util import TemplateVarRecorder, SortComponents
   ...: m = pe.ConcreteModel()
   ...: m.a = pe.Set(initialize=[1,2,3])
   ...: m.x = pe.Var(m.a)
   ...: e1 = m.x[1] * (m.x[2] + m.x[3])
   ...: e2 = m.x[1] - 1/m.x[2]
   ...: m.x[2].fix(2)
   ...: sorter = SortComponents.ORDERED_INDICES
   ...: vmap = {}
   ...: vr = TemplateVarRecorder(vmap, sorter)
   ...: visitor = LinearRepnVisitor(subexpression_cache={}, var_recorder=vr)
   ...: r1 = visitor.walk_expression(e1)

In [2]: r2 = visitor.walk_expression(e2)

In [3]: print(r2)
LinearRepn(mult=1, const=0, linear={140176910789184: 1}, nonlinear=- (1/x[2]))

Here, e2 should be linear because x[2] is fixed, but the LinearRepnVisitor says it is nonlinear. The problem is that the TemplateVarRecorder inserts variables in bulk based on the parent component. So, when the LinearRepnVisitor is used to walk e1, x[2] gets added to the var_map even though it is fixed. Then, LinearBeforeChildDispatcher._before_var does not behave as expected. With the variable fixed, we want to hit this line. However, because x[2] is in the var_map, we instead continue to this line, and the visitor thinks the variable is not fixed.

Note that this does not happen if we create the LinearRepnVisitor without the var_recorder.

Information on your system

Pyomo version: just pulled main

[cjck944084735-dot]

I can confirm this is still present on current main (your reporter-style reproduction gives LinearRepn(mult=1, const=-0.5, linear={x[1]: 1}, nonlinear=-(1/x[2])) today), and opened #4034: record_monomial now checks var.fixed before the var_map membership test, so variables registered in bulk while still free are correctly handled as constants once fixed. The regression tests pin down both the "registered while free, fixed later" path and the value substitution.